pith. sign in

arxiv: 2606.17941 · v1 · pith:WOVNYWN6new · submitted 2026-06-16 · 🧮 math.OC

PepsiCo Deploys AI-Driven Pricing and Promotion Optimization at Scale

Aleix Llenas , Eduardo Salazar-Trevi\~no , Francisco Leskovar , Francesc Pons-Llopis , Federico Todeschini , Deepika Gowda , David Bofill , Laxmi Anish
show 1 more author
Michael Cleavinger
This is my paper

Pith reviewed 2026-06-26 23:19 UTC · model grok-4.3

classification 🧮 math.OC
keywords pricing optimizationpromotion planningmixed-integer linear programmingBayesian hierarchical modelsmachine learning integrationrevenue managementnonlinear programmingenterprise optimization
0
0 comments X

The pith

Two optimization systems integrate machine learning forecasts with MILP and NLP solvers to automate pricing and promotion decisions at enterprise scale.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents two systems for revenue management that combine statistical learning with mathematical programming. One system pairs promotional forecasts from machine learning with a mixed-integer linear program to select calendars that maximize revenues while respecting constraints across channels. The second uses Bayesian hierarchical models to estimate elasticities and competitive effects, then applies nonlinear programming to recommend base prices over multiple periods. The central claim is that this integration produces scalable, automated decisions superior to manual methods when the forecasts and elasticity estimates hold. A reader would care because the approach addresses how operations research can handle high-dimensional planning problems in competitive markets where traditional methods fall short.

Core claim

The two systems couple machine learning-based promotional forecasts with a mixed-integer linear programming model to optimize promotional calendars across trade channels by searching millions of product-promotion-timing combinations, and use Bayesian hierarchical models for own- and cross-price elasticities and competitive interactions fed into a nonlinear programming engine for base price optimization, together demonstrating the feasibility and scalability of advanced optimization in large-scale enterprise environments for data-informed decisions aligned with strategic objectives.

What carries the argument

PromoAI and PricingAI, which integrate machine learning forecasts and Bayesian elasticity estimates with MILP and NLP solvers to search feasible plans and recommend decisions that maximize revenue and margin targets under customizable constraints.

If this is right

  • Automated search over millions of product-promotion-timing combinations produces revenue-maximizing calendars subject to business constraints.
  • Price recommendations account for demand elasticity, competitor actions, and financial targets across multi-period horizons.
  • The approach scales to large product and customer portfolios where manual or traditional methods become suboptimal.
  • Integration of statistical learning with programming enables automated, data-informed decisions aligned with strategic objectives.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same integration pattern could apply to other revenue management tasks such as assortment or inventory planning in similar industries.
  • Accuracy of the underlying forecasts and elasticity models would need ongoing validation against observed outcomes to sustain performance gains.
  • Real-time data streams could extend the systems to allow dynamic re-optimization as market conditions change.
  • Wider adoption might shift enterprise planning from periodic manual reviews to continuous model-driven processes.

Load-bearing premise

The machine-learning promotional forecasts and Bayesian hierarchical elasticity estimates are accurate enough representations of real demand responses and competitive interactions that feeding them into the MILP and NLP engines produces decisions superior to manual planning.

What would settle it

A controlled side-by-side comparison of actual revenue and margin results from the optimization systems versus manual planning over the same multi-month period and set of markets or product categories.

read the original abstract

Effective pricing and promotion planning constitutes a central pillar of strategic revenue management for firms operating in highly competitive and dynamic markets. These planning activities require the simultaneous consideration of demand elasticity, competitor actions, channel and market specific constraints, and financial objectives. As the dimensionality and interdependencies inherent in these problems increase, manual or traditional approaches become suboptimal and insufficient. In this context, Operations Research provides a robust methodological foundation for scalable, data-driven decision support systems that can optimize complex planning processes across large product and customer portfolios. This paper presents two large-scale optimization systems developed and deployed at PepsiCo to support Revenue Growth Management initiatives: PromoAI and PricingAI. PromoAI couples machine learning-based promotional forecasts with a mixed-integer linear programming model to optimize promotional calendars across trade channels, searching millions of product-promotion-timing combinations for the one that maximizes PepsiCo and retailer revenues subject to customizable business constraints. PricingAI optimizes base prices across product portfolios over multi-period horizons, using Bayesian hierarchical models to estimate own- and cross-price elasticities and competitive interactions, then feeding these into a nonlinear programming engine that recommends price changes aligned with revenue and margin targets under operational constraints. Together, these systems demonstrate the feasibility and scalability of advanced optimization in large-scale enterprise environments. They highlight the value of integrating statistical learning with mathematical programming to enable enterprise-level, automated decision-making that is both data-informed and aligned with strategic business objectives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript describes two deployed systems at PepsiCo for revenue growth management: PromoAI, which integrates machine-learning promotional forecasts with a mixed-integer linear program to optimize promotional calendars across channels by searching millions of combinations subject to business constraints; and PricingAI, which uses Bayesian hierarchical models for own- and cross-price elasticities and competitive effects, feeding these into a nonlinear program to set base prices over multi-period horizons under operational constraints. The central claim is that these systems demonstrate the feasibility and scalability of integrating statistical learning with mathematical programming for automated, data-informed enterprise decision-making.

Significance. If the described deployments are accurate, the paper supplies a concrete industrial case study of large-scale ML+OR integration in a major CPG firm, illustrating how optimization engines can handle high-dimensional promotional and pricing decisions under realistic constraints. This has illustrative value for practitioners and for the operations research community as an existence proof of enterprise-scale application, though the absence of model equations, validation statistics, or comparative performance data restricts its contribution to methodological advancement.

major comments (2)
  1. [Abstract] Abstract (paragraphs on PromoAI and PricingAI): the central feasibility and scalability claims rest on the assertion that the ML forecasts and elasticity estimates are sufficiently accurate to yield decisions superior to manual planning, yet no out-of-sample validation metrics, hold-out performance numbers, or constraint-handling details are supplied, preventing evaluation of whether the integrated systems actually achieve the claimed benefits.
  2. [Abstract] Abstract: the MILP and NLP engines are described only at the level of 'searching millions of combinations' and 'recommending price changes,' with no formulation of decision variables, objective functions, or the specific business constraints incorporated; this omission makes the scalability assertion impossible to assess within the mathematical optimization framework expected in this journal.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive comments on our manuscript describing the deployed PromoAI and PricingAI systems. We respond point-by-point to the major comments, clarifying the paper's scope as an industrial case study of large-scale deployment rather than a detailed methodological contribution.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraphs on PromoAI and PricingAI): the central feasibility and scalability claims rest on the assertion that the ML forecasts and elasticity estimates are sufficiently accurate to yield decisions superior to manual planning, yet no out-of-sample validation metrics, hold-out performance numbers, or constraint-handling details are supplied, preventing evaluation of whether the integrated systems actually achieve the claimed benefits.

    Authors: The manuscript is framed as a case study of enterprise deployment rather than a methods paper with benchmarked performance. Detailed out-of-sample metrics are not included because they are proprietary. We will revise the abstract to moderate the language, emphasizing successful operational deployment and integration of ML with optimization without claiming quantified superiority over manual planning. revision: partial

  2. Referee: [Abstract] Abstract: the MILP and NLP engines are described only at the level of 'searching millions of combinations' and 'recommending price changes,' with no formulation of decision variables, objective functions, or the specific business constraints incorporated; this omission makes the scalability assertion impossible to assess within the mathematical optimization framework expected in this journal.

    Authors: The high-level description reflects the paper's focus on demonstrating feasibility of integration at scale in a commercial setting. Full mathematical formulations, variables, objectives, and constraints are proprietary to PepsiCo and cannot be disclosed. This level of detail is consistent with other industrial application papers; we do not intend to expand the technical formulation. revision: no

standing simulated objections not resolved
  • Detailed out-of-sample validation metrics and full mathematical formulations of the MILP/NLP models cannot be provided due to commercial confidentiality.

Circularity Check

0 steps flagged

No significant circularity: descriptive deployment report

full rationale

The paper is a case study describing the deployment of PromoAI and PricingAI systems at PepsiCo. It presents no mathematical derivations, equations, fitted parameters, or predictions that could reduce to inputs by construction. The central claim—that the systems demonstrate feasibility and scalability of ML+optimization integration—is supported directly by the existence and scale of the reported deployments, with no load-bearing self-citation chains, ansatzes, or self-definitional steps. This is the most common honest finding for non-theoretical deployment reports.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is an industry case-study paper with no new mathematical models, derivations, or empirical claims that require free parameters, axioms, or invented entities; all referenced techniques are standard in the operations-research literature.

pith-pipeline@v0.9.1-grok · 5819 in / 1156 out tokens · 35154 ms · 2026-06-26T23:19:26.492957+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 2 canonical work pages

  1. [1]

    1958 , publisher=

    The Evolution of Intelligence: The Nervous System as a Model of Its Environment , author=. 1958 , publisher=

    1958

  2. [2]

    Journal of Machine Learning Research , volume=

    Automatic differentiation variational inference , author=. Journal of Machine Learning Research , volume=

  3. [3]

    Interfaces , volume=

    Marriott International increases revenue by implementing a group pricing optimizer , author=. Interfaces , volume=. 2010 , publisher=

    2010

  4. [4]

    2023 , publisher=

    Alibaba realizes millions in cost savings through integrated demand forecasting, inventory management, price optimization, and product recommendations , author=. 2023 , publisher=

    2023

  5. [5]

    Interfaces , volume=

    Harvest hope food bank optimizes its promotional strategy to raise donations using integer programming , author=. Interfaces , volume=. 2018 , publisher=

    2018

  6. [6]

    Journal Of Statistical Software , volume=

    Stan: A probabilistic programming language , author=. Journal Of Statistical Software , volume=. 2017 , publisher=

    2017

  7. [7]

    and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and Haberland, Matt and Reddy, Tyler and Cournapeau, David and Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and Bright, Jonathan and. Nature Methods , year =

  8. [8]

    AIChE Journal , volume=

    Piecewise MILP under-and overestimators for global optimization of bilinear programs , author=. AIChE Journal , volume=. 2008 , publisher=

    2008

  9. [9]

    Why AI Transformations Should Start with Pricing , journal =

    Jo. Why AI Transformations Should Start with Pricing , journal =. 2021 , url =

    2021

  10. [10]

    MathCo , year =

    Silvana Dimitrov and Aditya Durai , title =. MathCo , year =

  11. [11]

    PepsiCo , year = 2025, journal=

    2025

  12. [12]

    Management Science , volume=

    From predictive to prescriptive analytics , author=. Management Science , volume=. 2020 , publisher=

    2020

  13. [13]

    1999 , publisher=

    Integer and combinatorial optimization , author=. 1999 , publisher=

    1999

  14. [14]

    1997 , publisher=

    Introduction to linear optimization , author=. 1997 , publisher=

    1997

  15. [15]

    Annals Of Statistics , pages=

    Greedy function approximation: a gradient boosting machine , author=. Annals Of Statistics , pages=. 2001 , publisher=

    2001

  16. [16]

    Journal of the Royal Statistical Society Series A: Statistics in Society , volume =

    Hewson, Paul , title =. Journal of the Royal Statistical Society Series A: Statistics in Society , volume =. 2015 , month =. doi:10.1111/j.1467-985X.2014.12096_1.x , url =

    work page doi:10.1111/j.1467-985x.2014.12096_1.x 2015

  17. [17]

    Operations Research , volume=

    Online network revenue management using thompson sampling , author=. Operations Research , volume=. 2018 , publisher=

    2018

  18. [18]

    Production and Operations Management , volume=

    Dynamic pricing through data sampling , author=. Production and Operations Management , volume=. 2018 , publisher=

    2018

  19. [19]

    Data-Driven Optimization in Revenue Management: Pricing, Assortment Planning, and Demand Learning , author=

  20. [20]

    Management Science , volume=

    Dynamic pricing (and assortment) under a static calendar , author=. Management Science , volume=. 2021 , publisher=

    2021

  21. [21]

    Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

    Optimization beyond prediction: Prescriptive price optimization , author=. Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

  22. [22]

    Advances in Neural Information Processing Systems , pages=

    LightGBM: A Highly Efficient Gradient Boosting Decision Tree , author=. Advances in Neural Information Processing Systems , pages=

  23. [23]

    Journal of Global Optimization , volume=

    Differential evolution--a simple and efficient heuristic for global optimization over continuous spaces , author=. Journal of Global Optimization , volume=. 1997 , publisher=

    1997

  24. [24]

    2006 , publisher=

    The theory and practice of revenue management , author=. 2006 , publisher=

    2006

  25. [25]

    Johns Hopkins Carey Business School Research Paper , number=

    Customer-driven bundle promotion optimization at scale , author=. Johns Hopkins Carey Business School Research Paper , number=

  26. [26]

    Management Science 66(3):1025--1044

    Bertsimas D, Kallus N (2020) From predictive to prescriptive analytics. Management Science 66(3):1025--1044

    2020

  27. [27]

    Bertsimas D, Tsitsiklis JN (1997) Introduction to linear optimization, volume 6 (Athena Scientific Belmont, MA)

    1997

  28. [28]

    Journal of Statistical Software 76(1):1--32

    Carpenter B, Gelman A, Hoffman MD, Lee D, Goodrich B, Betancourt M, Brubaker M, Guo J, Li P, Riddell A (2017) Stan: A probabilistic programming language. Journal of Statistical Software 76(1):1--32

    2017

  29. [29]

    Production and Operations Management 27(6):1074--1088

    Cohen MC, Lobel R, Perakis G (2018) Dynamic pricing through data sampling. Production and Operations Management 27(6):1074--1088

    2018

  30. [30]

    INFORMS Journal on Applied Analytics 53(1):32--46

    Deng Y, Zhang X, Wang T, Wang L, Zhang Y, Wang X, Zhao S, Qi Y, Yang G, Peng X (2023) Alibaba realizes millions in cost savings through integrated demand forecasting, inventory management, price optimization, and product recommendations. INFORMS Journal on Applied Analytics 53(1):32--46

    2023

  31. [31]

    MathCo ://mathco.com/article/digitizing-rgm-for-scale-and-impact/

    Dimitrov S, Durai A (2023) Digitizing revenue growth management for scale and impact. MathCo ://mathco.com/article/digitizing-rgm-for-scale-and-impact/

    2023

  32. [32]

    Johns Hopkins Carey Business School Research Paper (22-14)

    Fattahi A, Li Y, Sahin O (2022) Customer-driven bundle promotion optimization at scale. Johns Hopkins Carey Business School Research Paper (22-14)

    2022

  33. [33]

    Operations Research 66(6):1586--1602

    Ferreira KJ, Simchi-Levi D, Wang H (2018) Online network revenue management using thompson sampling. Operations Research 66(6):1586--1602

    2018

  34. [34]

    Annals Of Statistics 1189--1232

    Friedman JH (2001) Greedy function approximation: a gradient boosting machine. Annals Of Statistics 1189--1232

    2001

  35. [35]

    Gelman A, Carlin JB, Stern HS, Rubin DB (1995) Bayesian data analysis (Chapman and Hall/CRC)

    1995

  36. [36]

    ://www.gurobi.com

    Gurobi Optimization, LLC (2024) Gurobi Optimizer Reference Manual . ://www.gurobi.com

    2024

  37. [37]

    Boston Consulting Group ://www.bcg.com/publications/2021/ai-pricing-tranformations

    Hazan J, Br \'e g \'e C, Verwaerde JS, Bassoulet A (2021) Why ai transformations should start with pricing. Boston Consulting Group ://www.bcg.com/publications/2021/ai-pricing-tranformations

    2021

  38. [38]

    Interfaces 40(1):47--57

    Hormby S, Morrison J, Dave P, Meyers M, Tenca T (2010) Marriott international increases revenue by implementing a group pricing optimizer. Interfaces 40(1):47--57

    2010

  39. [39]

    Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 1833--1841

    Ito S, Fujimaki R (2017) Optimization beyond prediction: Prescriptive price optimization. Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 1833--1841

    2017

  40. [40]

    Advances in Neural Information Processing Systems, 3146--3154

    Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, Ye Q, Liu TY (2017) Lightgbm: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 3146--3154

    2017

  41. [41]

    Journal of Machine Learning Research 18(14):1--45

    Kucukelbir A, Tran D, Ranganath R, Gelman A, Blei DM (2017) Automatic differentiation variational inference. Journal of Machine Learning Research 18(14):1--45

    2017

  42. [42]

    Management Science 67(4):2292--2313

    Ma W, Simchi-Levi D, Zhao J (2021) Dynamic pricing (and assortment) under a static calendar. Management Science 67(4):2292--2313

    2021

  43. [43]

    Miao S (2020) Data-Driven Optimization in Revenue Management: Pricing, Assortment Planning, and Demand Learning. Ph.D. thesis

    2020

  44. [44]

    PepsiCo (2025) ://www.pepsico.com/who-we-are/about-pepsico

    2025

  45. [45]

    Journal of Global Optimization 11(4):341--359

    Storn R, Price K (1997) Differential evolution--a simple and efficient heuristic for global optimization over continuous spaces. Journal of Global Optimization 11(4):341--359

    1997

  46. [46]

    Talluri KT, Van Ryzin GJ (2006) The theory and practice of revenue management, volume 68 (Springer Science & Business Media)

    2006

  47. [47]

    E., et al

    Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau D, Burovski E, Peterson P, Weckesser W, Bright J, van der Walt SJ, Brett M, Wilson J, Millman KJ, Mayorov N, Nelson ARJ, Jones E, Kern R, Larson E, Carey CJ, Polat \.I , Feng Y, Moore EW, VanderPlas J, Laxalde D, Perktold J, Cimrman R, Henriksen I, Quintero EA, Harris CR, Archibald AM, R...

    work page doi:10.1038/s41592-019-0686-2 2020

  48. [48]

    AIChE Journal 54(4):991--1008

    Wicaksono DS, Karimi IA (2008) Piecewise MILP under- and overestimators for global optimization of bilinear programs. AIChE Journal 54(4):991--1008

    2008

  49. [49]

    Wolsey LA, Nemhauser GL (1999) Integer and combinatorial optimization (John Wiley & Sons)

    1999